transcranial
resnet top-5 classes showing in Opera, not Firefox
First, very cool - great work! :-D
For your reference, keras-js runs fast in Opera; appears to run (slower) in Firefox, but does not display the top-5 classes (50-layer ResNet tried; supplied wolf image).
Firefox 49.0.1 - lots of plugins [re: this issue, disabled AdBlockPlus; uBlock Origin; Ghostery] Opera 40.0 - minimal plugins Arch Linux x86_64; Intel Core i7-4790 CPU; ...
Private message me -- if desired -- for my Firefox addons, configuration [about:support]; I don't want to share them, publicly.
Edit - updated info:
The website
https://transcranial.github.io/keras-js/
will not run (load) in Firefox started in Safe Mode (add-ons disabled; ...). [I wanted to see if disabling add-ons helped viz-a-viz the ResNet implementation in Firefox.]
Wolf image: https://upload.wikimedia.org/wikipedia/commons/0/00/Yellowstone-wolf-17120.jpg
Inception-v3 [ImageNet]:
-
Firefox:
dogsled 53% apiary 8% seashore 3 jigsaw puzzle 3% dowitcher 2%
-
Opera:
cairn 74% silky terrier 6% Yorkshire terrier 5% Australian terrier 5% Norwich terrier 3%
50-Layer ResNet:
-
Firefox:
0% 0% 0% 0% 0%
-
Opera:
timber wolf 92% coyote 3% dogsled 1% white wolf 1% red wolf 1%
@victoriastuart interesting, so the resnet is problematic in firefox but inception-v3 is fine? Thanks for the report, will take a look. There are some issues with firefox I have to look into in any case.
Yes, exactly. Perhaps this is related: keras-js will load/process this wolf image
http://media.salon.com/2015/07/wolf.jpg
in those two networks { Inception | resnet } in Opera, but not Firefox: "Error loading URL." So, e.g., what we humans see and what Opera vs. Firefox returns to keras-js may be different: perhaps browser-specific image format/metadata? That might also explain the dramatic differences in classifications in the classifications of my first wolf image in those two nets, in Firefox? Just guessing here, however.
Again, great work: I'm sure you'll have the issues resolved in no time! ;-)